Deep Sketch Vectorization in the browser

A browser port of Deep Sketch Vectorization via Implicit Surface Extraction (Chuan Yan, Yong Li, Deepali Aneja, Matthew Fisher, Edgar Simo-Serra, Yotam Gingold; SIGGRAPH 2024). Raster sketches are vectorized entirely on your device: the networks run with ONNX Runtime Web (WebGPU, or WebAssembly as a fallback), and nothing is uploaded.

Try it: https://yig.github.io/Deep-Sketch-Vectorization-Web/

Features:

• the automatic pipeline of the reference implementation (light and full models), verified stage by stage against it;
• interactive refinement: edit keypoints and the under-sampling map, with undo and saved edits;
• optional pre-processing with Edgar Simo-Serra's line thinning and Xiaoyu Xiang's Anime2Sketch;
• RDP simplification or Bézier fitting, and SVG download.

Development

The ONNX models are generated from the pretrained PyTorch weights and are not in git. With Deep-Sketch-Vectorization and the weights checked out next to this repository (as ../Deep-Sketch-Vectorization and ../pretrained):

npm install
npm run models   # exports public/models/*.onnx (needs uv)
npm run dev

npm test runs the unit tests; the golden tests need fixtures generated by tools/make_fixtures.py and tools/make_preprocess_fixtures.py. See DESIGN.md for the design and architecture.

Pushing to main builds the site and deploys it to GitHub Pages (.github/workflows/deploy.yml). Visits and vectorized sketches are counted anonymously with GoatCounter (src/analytics.ts).